Human Detection, Tracking and Segmentation in Surveillance Video

Explore human detection, tracking, and segmentation techniques in surveillance video through this doctoral dissertation defense. Delve into scene-specific learning approaches, including DPM human detectors, superpixel-based Bag-of-Words classifiers, and part-based person-specific SVM models. Discover methods for handling occlusions in detection and tracking, as well as separating human and background superpixels using Conditional Random Fields. Learn about leveraging spatio-temporal constraints with tracklet-based Gaussian Mixture Models and multi-frame graph optimization. Examine the development of NONA, an efficient real-time tracking system for high-definition surveillance video, implemented using Intel Threading Building Blocks. Gain insights into Fast Fourier Transform-based normalized cross-correlation, Adaptive Template scaling, and Local Frame Differencing techniques for improved tracking performance.

Motivation
Video Surveillance Tasks
Outline
Problems
Initial Detection
Training and Classification
Iteratively Learning
Superpixel Segmentation
Bag-of-Words
Qualitative Results
From Detection to Tracking
Part-based Model in Tracking
Features and Classifiers
Data Association
Proposed Method
DPM with Occlusion Handling
Occlusion handling Results
Occlusion Handling in Tracking
Occlusion Reasoning Results
Quantitative Results -- Town Center
Boston Airport
Parking Lot 1
Parking Lot dataset
From Detection to Segmentation
Human Detection
Background Gaussian Mixture Model (GMM)
Part-based Detection Potential
Graph Optimization
Initial Results
Multi-frame Segmentation
Obtaining Tracklets
Multi-frame CRF Optimization
Datasets and Groundtruth
Comparison with Background Subtraction
Segmentation Results (by frame)
For Real-World Application
Objective: Tracking
Multi-threaded Implementation
Tracking Overview
Adaptive Scaling
Local Frame Differencing
Summary
Dissertation Conclusion
Future Work
Publication